DFMEA analysis method and computer program product

Abstract

The invention provides a DFMEA analysis method and a computer program product, and the method comprises the steps: firstly determining an analysis main body and an analysis intention based on a failure question after obtaining the failure question related to a to-be-tested product; based on the analysis main body and the failure questions, a keyword query mode and a vector query mode are adopted to retrieve various reference analysis knowledge from a pre-constructed multi-source knowledge base; and finally, based on the analysis subject, the analysis intention and all reference analysis knowledge, calling a large language model, and generating a DFMEA analysis result corresponding to the analysis subject. The multi-source knowledge base is created on the basis of various heterogeneous data sources and comprises a large amount of analysis knowledge, so that reference analysis knowledge can be accurately searched from the multi-source knowledge base by adopting a keyword query mode and a vector query mode, and finally, a large language model is called to generate a DFMEA analysis result, so that DFMEA analysis is automatically performed, and the efficiency is improved. Manual analysis is not needed, and the analysis efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, and more specifically, to a DFMEA analysis method and computer program product. Background Technology

[0002] Design Failure Mode and Effects Analysis (DFMEA) is a key engineering method for identifying and mitigating potential design risks in the early stages of product development, ensuring product reliability and safety.

[0003] In existing technologies, the implementation of DFMEA mainly relies on the personal experience and knowledge reserves of engineers for analysis. This results in highly subjective analysis results, difficulty in reusing knowledge, and is very time-consuming and labor-intensive, leading to low analysis efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a DFMEA analysis method and computer program product to improve the problems existing in the prior art.

[0005] The embodiments of the present invention can be implemented as follows: In a first aspect, the present invention provides a DFMEA analysis method, comprising: Obtain failure questions related to the product under test; The analysis subject and analysis intent are determined based on the aforementioned failure questions; Based on the analysis subject and the failure query, various reference analysis knowledge is retrieved from a pre-built multi-source knowledge base using keyword query and vector query methods, respectively; the multi-source knowledge base is created based on various heterogeneous data sources and includes a large amount of analysis knowledge. Based on the analysis subject, the analysis intent, and all the reference analysis knowledge, the large language model is invoked to generate the DFMEA analysis results corresponding to the analysis subject.

[0006] In an optional implementation, the multi-source knowledge base is constructed in the following manner: Extract the core field data of each table item from the DFMEA table in the DFMEA database; The core field data is vectorized using a vectorization model to obtain the semantic vector corresponding to each core field data. For any unstructured document obtained, based on the single input limitation of the vectorization model, the unstructured document is segmented into multiple semantically complete paragraphs, and each paragraph is converted into a corresponding semantic vector using the vectorization model. For any acquired design drawing, a multimodal model is invoked to extract visual feature vectors and text descriptions from the design drawing; For any acquired complaint recording, based on the complaint recording, extract the key complaint content related to the product failure and its timestamp, and use the vectorization model to convert the key complaint content into a corresponding semantic vector; For any acquired maintenance video, key graphic information is extracted based on the maintenance video, and the key graphic information is converted into a corresponding comprehensive vector; Each core field data and its corresponding semantic vector, each paragraph in the unstructured document and its corresponding semantic vector, the design drawings and their visual feature vectors and text descriptions, the complaint recordings and their key complaint content and semantic vectors, and the maintenance videos and their key graphic and textual information and comprehensive vectors are associated and stored to obtain the multi-source knowledge base.

[0007] In an optional implementation, the step of extracting key graphic information based on the maintenance video and converting the key graphic information into a corresponding comprehensive vector includes: Extract the video narration and video content description from the maintenance video; Identify key speech content related to fault analysis from the video speech content; Based on the timestamp of the key spoken content in the maintenance video, at least one keyframe image is extracted from the maintenance video; the key graphic information includes the video content description, the key spoken content, and all keyframe images; The video content description and key spoken content are converted into text content vectors using the vectorization model, and visual content vectors are extracted from all keyframe images using the multimodal model. The text content vector and the visual content vector are weighted and fused to obtain the comprehensive vector.

[0008] In an optional implementation, the multi-source knowledge base includes a relational database and a vector database. The relational database includes several analytical knowledge items, their source categories, and inclusion times. The vector database includes vectors corresponding to each type of analytical knowledge item. The steps of retrieving various reference analysis knowledge from a pre-built multi-source knowledge base based on the analysis subject and the failure query, using keyword query and vector query methods respectively, include: Search for each synonym that matches the analyzed subject from a preset thesaurus; For each type of analytical knowledge, a preset keyword matching algorithm is used to calculate the matching degree between the analytical knowledge and each subject. Then, the mean of each matching degree is calculated to obtain the lexical relevance of the analytical knowledge. The subject is the analytical subject or the synonymous subject. The failure question is converted into a question vector using a vectorization model, and the cosine distance between the question vector and the vector corresponding to each type of analytical knowledge is calculated to obtain the semantic relevance of each type of analytical knowledge. The lexical relevance and semantic relevance corresponding to each type of analytical knowledge are weighted and summed to obtain the comprehensive relevance corresponding to each type of analytical knowledge; Based on the comprehensive relevance, source category, and inclusion time of each type of analytical knowledge, all analytical knowledge is reordered to filter multiple types of reference analytical knowledge.

[0009] In an optional implementation, the step of reordering all the analytical knowledge based on the comprehensive relevance, source category, and inclusion time corresponding to each type of analytical knowledge to filter multiple types of reference analytical knowledge includes: For each type of analytical knowledge, a source authority score matching the source category of the analytical knowledge is determined from the mapping relationship between multiple source categories and multiple preset scores; Input the difference between the time the analytical knowledge was collected and the current time into a preset decay model to obtain the timeliness score corresponding to the analytical knowledge; The comprehensive score of the analytical knowledge is obtained by weighting and summing the comprehensive relevance, source authority score, and failure score corresponding to the analytical knowledge. After arranging all analytical knowledge in descending order of the comprehensive score, the top K analytical knowledge items are used as the reference analytical knowledge items.

[0010] In an optional implementation, the step of generating the DFMEA analysis result corresponding to the analysis subject by calling a large language model based on the analysis subject, the analysis intent, and all the reference analysis knowledge includes: The analysis subject, the analysis intent, all the reference analysis knowledge, and the preset task instructions are combined into a Prompt text. The Prompt text is input into the large language model to obtain the DFMEA analysis results corresponding to the analysis subject. The Prompt text is used to instruct the large language model to summarize and generate multiple DFMEA entries related to the potential failure consequences of the analysis subject based on the input content, and output each DFMEA entry in a preset format after providing a brief description.

[0011] In an optional implementation, the DFMEA analysis results include multiple DFMEA entries, and the method further includes: Obtain the review comments from verification experts for each of the aforementioned DFMEA entries; If the review opinion for any of the DFMEA entries is to directly adopt it, then the current verification time and the DFMEA entry will be stored in the DFMEA database; If the review opinion of any of the DFMEA entries is that the modification is adopted, then the current verification time and the modified DFMEA entry are stored in the DFMEA database, and the Prompt text, the DFMEA entry and the modified DFMEA entry are combined into a sample data and stored in the sample database. If the review opinion of any of the DFMEA entries is rejection, the reason for rejection of the DFMEA entry is obtained, and the DFMEA entry and the reason for rejection are stored in the optimization library; the optimization library is used to guide technical personnel to adjust the prompt words; From the DFMEA analysis results, delete each DFMEA entry whose review comment is "rejected", and replace each DFMEA entry whose review comment is "accepted after modification" with the corresponding modified DFMEA entry to obtain the target DFMEA analysis results.

[0012] In an optional implementation, the method further includes: If the number of samples in the sample library reaches a preset value, then the large language model is post-trained based on all the sample data in the sample library using a contrastive learning approach.

[0013] In an optional implementation, the method further includes: Obtain the target DFMEA analysis results corresponding to multiple analysis subjects of the product under test; Based on preset prompts and the target DFMEA analysis results corresponding to multiple analysis subjects, a large language model is invoked to generate a DFMEA analysis report for the product under test. The preset prompts are used to instruct the large language model to integrate and analyze the DFMEA analysis results of each target according to the hierarchical relationship of multiple analysis subjects in order to generate the DFMEA analysis report.

[0014] In a second aspect, the present invention provides a computer program product that, when executed by a processor, implements the method described in the first aspect above.

[0015] Compared with existing technologies, this invention provides a DFMEA analysis method and computer program product. After obtaining failure questions related to the product under test, the analysis subject and analysis intent are first determined based on the failure questions. Based on the analysis subject and failure questions, various reference analysis knowledge is retrieved from a pre-built multi-source knowledge base using keyword query and vector query methods, respectively. Finally, based on the analysis subject, analysis intent, and all reference analysis knowledge, a large language model is invoked to generate the DFMEA analysis result corresponding to the analysis subject. The multi-source knowledge base of this invention is created based on multiple heterogeneous data sources and includes a large amount of analysis knowledge. Therefore, the keyword query and vector query methods can accurately retrieve reference analysis knowledge from the multi-source knowledge base, and the final DFMEA analysis result is generated by invoking a large language model. This automates the DFMEA analysis, eliminating the need for manual analysis and improving analysis efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is one of the flowcharts of a DFMEA analysis method provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram illustrating the construction process of a multi-source knowledge base provided in an embodiment of the present invention.

[0019] Figure 3 This is a second schematic flowchart of a DFMEA analysis method provided in an embodiment of the present invention.

[0020] Figure 4 This is the third flowchart illustrating a DFMEA analysis method provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0024] It should be noted that, where there is no conflict, the features in the embodiments of the present invention can be combined with each other.

[0025] In existing technologies, the implementation of DFMEA mainly suffers from the following pain points: (1) Heavy reliance on expert experience: DFMEA analysis workshops are usually conducted in the form of brainstorming, and the depth and breadth of the analysis are highly dependent on the personal experience and knowledge reserves of the participating engineers. This leads to highly subjective analysis results, difficulty in reusing knowledge, and the risk of loss of core knowledge due to personnel changes.

[0026] (2) Data silos and low knowledge utilization: Enterprises contain a large amount of data related to product failures, such as historical DFMEA reports, product manuals, after-sales maintenance records, 3D design drawings, user complaint recordings, and even maintenance videos of equipment failures. However, these data are in various formats (structured, text, images, audio and video, etc.) and are scattered across different systems, forming "data silos". Traditional analysis methods cannot effectively integrate and utilize this multimodal unstructured data, resulting in a large amount of valuable tacit knowledge being buried.

[0027] (3) The analysis is inefficient and incomplete: The process of manually reviewing materials, organizing meetings and discussions, and compiling reports is time-consuming and laborious. At the same time, due to the limitations of human cognition and memory, the analysis process is prone to missing some occasional, atypical or cross-domain failure modes, resulting in incomplete risk coverage.

[0028] Most existing knowledge bases or auxiliary software systems only support searching structured tabular data or keywords, failing to deeply understand the semantics of unstructured text, and lacking the ability to process and understand non-textual data such as images and videos, thus failing to fundamentally solve the aforementioned problems. Therefore, there is an urgent need for a new technological solution that can break down data barriers, deeply understand multimodal information, and intelligently generate analytical entries.

[0029] In view of this, embodiments of the present invention provide a DFMEA analysis method that can integrate multiple heterogeneous data sources related to product failure into a multi-source knowledge base. Therefore, for failure queries of the product under test, keyword query and vector query methods can be used to accurately find reference analysis knowledge from the multi-source knowledge base, and finally call the large language model to generate DFMEA analysis results. This automates DFMEA analysis without the need for manual analysis, thus improving analysis efficiency.

[0030] The product under test mentioned in this invention can be any product that may have safety, functional, compliance, or user satisfaction issues due to design flaws. These products can be identified and their potential failure risks reduced through DFMEA analysis. For example, the product under test can be, but is not limited to, various products in the automotive industry, various products in the medical device industry, various products in the aerospace industry, and various electrical appliances in the home appliance industry.

[0031] The following uses a kitchen appliance as an example to illustrate the DFMEA analysis method provided in this embodiment of the invention.

[0032] Please see Figure 1 , Figure 1 This is a flowchart illustrating a DFMEA analysis method provided in an embodiment of the present invention. The execution subject of this method can be a computing device such as a smartphone, personal laptop, personal computer, or server. The method includes the following steps S201 to S204.

[0033] S201. Obtain failure questions related to the product under test.

[0034] DFMEA analysis may involve different levels and aspects of the product to be purchased, and the failure question can be raised by the engineer for one aspect of one level.

[0035] S202. Determine the subject and intent of analysis based on failure questions.

[0036] In this embodiment, the timeliness question can be identified by intent recognition to determine the analysis subject and analysis intent. The analysis subject refers to the analysis object targeted by the failure question.

[0037] For example, suppose the failure query is "What are the common failure consequences of thermocouples in stoves?", then the subject of analysis is "the thermocouples in stoves", and the analysis intent is "to query failure consequences". It should be noted that this example is only for illustration and is not intended to be limiting.

[0038] S203. Based on the analysis subject and failure question, various reference analysis knowledge is retrieved from the pre-constructed multi-source knowledge base using keyword query and vector query methods respectively.

[0039] In this embodiment, the multi-source knowledge base is a knowledge base created based on multiple heterogeneous data sources, including a large amount of analytical knowledge. Furthermore, by employing a hybrid retrieval method using keyword search and vector search, it can accurately find effective reference analytical knowledge.

[0040] S204. Based on the analysis subject, analysis intent, and all reference analysis knowledge, call the large language model to generate the DFMEA analysis results corresponding to the analysis subject.

[0041] In this embodiment, the analysis subject, analysis intent, and all reference analysis knowledge are used as prior knowledge. A large language model is invoked to generate the DFMEA analysis result corresponding to the current analysis subject. This enables knowledge integration and avoids model illusion. The final DFMEA analysis result may include multiple DFMEA entries, which may include potential failure consequences determined based on a reference analysis, a brief description, and the source of evidence.

[0042] Optionally, multi-dimensional, cross-modal heterogeneous data sources can include structured DFMEA databases, unstructured documents (such as product manuals, failure analysis cases, technical reports, user manuals, or expert experience summaries), product or component design drawings, user complaint recordings, and maintenance videos. These heterogeneous data are generally stored in different systems, making it difficult to achieve effective retrieval and correlation analysis at a unified semantic level. This invention systematically integrates information from different data sources and uses vectorization and multimodal models to process text, visual, and mixed-modal content respectively, thereby achieving semantic alignment and structured organization of heterogeneous data within a unified knowledge space.

[0043] That is, please see Figure 2 The process of constructing a multi-source database may include the following steps S101 to S107.

[0044] S101. Extract the core field data of each table item from the DFMEA table in the DFMEA database.

[0045] S102. Use a vectorization model to vectorize each core field data to obtain the semantic vector corresponding to each core field data.

[0046] The DFMEA tables in the DFMEA database are typically completed by R&D engineers in the early stages of product development according to industry standards (such as the AIAG-VDA FMEA manual). They cover multi-dimensional technical items, from functional definition to risk control. Each item in the DFMEA table represents an independent analysis node, and its fields may include, but are not limited to: name, functional description, potential failure mode, failure consequences, failure causes, control measures, failure frequency, detectability score, risk priority number (RPN), recommended improvement measures, and implementation status of the measures. These fields together constitute a structured expression of the potential technical defects and their impact on a specific design stage.

[0047] In this embodiment, each core field data extracted from the DFMEA table may include: name, functional description, potential failure mode, failure consequences, failure cause, and control measures. Then, a vectorization model is used to convert each core field data into a semantic vector.

[0048] The vectorization model can be a pre-trained language model based on a deep neural network architecture. Through self-supervised learning on large-scale text corpora, it acquires general language representation capabilities. The role of the vectorization model is to map non-numerical text information to a high-dimensional continuous vector space, making semantically similar content appear closer together in the vector space. For example, this vectorization model may include, but is not limited to, the BERT (Bidirectional Encoder Representations from Transformers) model and its optimized variants, such as the RoBERTa (Robustly optimized BERT pretraining approach) model, the DistilBERT model, or the Sentence-BERT model. It should be noted that this example is merely illustrative and not intended to be limiting.

[0049] S103. For any unstructured document obtained, based on the single input limitation of the vectorization model, the unstructured document is divided into multiple semantically complete paragraphs, and each paragraph is converted into a corresponding semantic vector using the vectorization model.

[0050] In this embodiment, an unstructured document that is valuable for failure analysis often exceeds the single-input limit (i.e., single-input token limit) of the vectorization model. Therefore, it is necessary to first segment the unstructured document into multiple semantically complete paragraphs based on the single-input limit of the vectorization model using the semantic chunking method. The segmentation strategy can be implemented based on sentence boundaries, paragraph structure, or semantic coherence algorithms to ensure that each segment has semantic independence and integrity. Subsequently, each paragraph is converted into a corresponding semantic vector using the same vectorization model, forming a "paragraph-vector" mapping pair.

[0051] S104. For any of the acquired design drawings, call the multimodal model to extract visual feature vectors and text descriptions from the design drawings.

[0052] In this embodiment, a design drawing that is valuable for failure analysis is used for joint feature extraction because it contains multimodal information such as geometric figures and labeled text. This multimodal model can be a Vision-Language Pre-training Model (VLP), such as CLIP (Contrastive Language–Image Pretraining) or a similar architecture, capable of simultaneously parsing visual elements and accompanying text descriptions in an image.

[0053] In this process, the output of the multimodal model includes two parts: one is the visual feature vector extracted from the design drawings, which reflects its topology, component layout and graphic style; the other is the text description generated from the text area in the drawing, which is used to explain the functional modules or assembly relationships.

[0054] For example, if the design drawing is an assembly drawing of a stove valve body, then the extracted text description could be "Stove valve body assembly drawing, including thermocouple fixing ports". It should be noted that this example is merely illustrative and not intended to be limiting.

[0055] S105. For any acquired complaint recording, extract the key complaint content related to the product malfunction and its timestamp based on the complaint recording, and use a vectorization model to convert the key complaint content into the corresponding semantic vector.

[0056] In this embodiment, the complaint recordings typically document user feedback on product malfunctions encountered during actual use. To extract valid fault clues from these recordings, ASR (Automatic Speech Recognition) technology is used to transcribe the complaint recordings into timestamped text. Furthermore, keyword detection, intent recognition, or event extraction methods are employed to locate key complaint content related to the product malfunction and its timestamp from the text content. Key complaint content may include specific descriptions of malfunctions such as "black screen upon startup," "slow performance," or "unresponsive buttons." Subsequently, the aforementioned vectorization model is used to convert this key complaint content into a corresponding semantic vector, enabling it to be compared and correlated with other textual knowledge in the semantic space.

[0057] For example, consider a complaint recording: complaint_call_345.wav, whose text includes: "...My stove, every time I light it I have to hold down the switch for a long time, and the flame goes out as soon as I release it! It's so annoying..." The key complaint is "My stove, every time I light it I have to hold down the switch for a long time, and the flame goes out as soon as I release it." It should be noted that this example is merely illustrative and is not intended to be limiting.

[0058] S106. For any acquired maintenance video, extract key graphic and textual information based on the maintenance video, and convert the key graphic and textual information into a corresponding comprehensive vector.

[0059] In this embodiment, key graphic information may include video content description, key spoken content, and key frame images. The key spoken content is a description of the fault phenomenon or fault analysis, while the key frame images mainly include video frames of components or products that have faults or abnormalities.

[0060] Optionally, step S106 may be implemented in ways that include S1061 to S1065.

[0061] S1061. Extract video narration and video content description from maintenance videos.

[0062] In this embodiment, ASR technology can be used to extract video speech content from the maintenance video, or the aforementioned large language model supporting multimodal input can be directly used to extract video speech content and video content description based on the maintenance video. The video content description is a summary of the maintenance video, which may include descriptions of the repair actions (such as "disassembly" and "cleaning") performed by the maintenance personnel on the repair object (such as "thermocouple") and fault descriptions (such as "blackening of the tip"), for example, "The maintenance personnel are disassembling the stove thermocouple," or "Carbon deposits were detected on the tip of the component."

[0063] S1062. Identify key speech content related to fault analysis from the video speech content.

[0064] In this embodiment, the extracted video speech content is semantically filtered to identify language segments that are directly related to product malfunction phenomena, failure cause judgment, or maintenance measures. These selected parts constitute the key speech content.

[0065] For example, key points of the speech could be, "The problem of slow ignition is because the thermocouple signal is too weak," or "Look, this head is all burnt black, that's carbon buildup. The thermocouple isn't aligned with the burner; just replace it."

[0066] S1063. Based on the timestamps of key spoken content in the maintenance video, extract at least one keyframe image from the maintenance video.

[0067] In this embodiment, to establish the spatiotemporal correlation between voice and visual information, at least one keyframe image is extracted from the maintenance video based on the timestamps of key spoken content in the video. These keyframes typically correspond to core nodes of the maintenance operation or scenes where fault characteristics are most obvious, such as close-ups of damaged components or scenes of measurements being performed. For example, a keyframe image may include a blackened thermocouple tip.

[0068] Among these methods, keyframe extraction and target detection techniques can be used to extract at least one keyframe image from the maintenance video.

[0069] S1064. Use a vectorization model to convert video content descriptions and key spoken content into text content vectors, and use a multimodal model to extract visual content vectors from all keyframe images.

[0070] In this embodiment, a vectorization model is used to convert video content descriptions and key spoken content into text content vectors. This process maps the two types of text information to a unified semantic space, forming a vectorized representation of the language expression in the video. At the same time, a multimodal model is used to extract visual content vectors from all keyframe images. This model captures visual elements such as spatial structure, texture patterns, and target shapes by performing deep feature encoding on the images.

[0071] S1065. Weighted fusion of text content vector and visual content vector to obtain comprehensive vector.

[0072] The text content vector and the visual content vector are weighted and fused to obtain a comprehensive vector. This fusion process coordinates the information contributions of text and images by setting weight coefficients, so that the generated comprehensive vector can simultaneously reflect the semantic statements and visual evidence in the maintenance video, forming a compact and discriminative overall representation of the original video content.

[0073] Through steps S1061-S1065, the maintenance video can be analyzed using both audio and visual channels, and image-text alignment can be achieved by combining timestamps. This effectively integrates multimodal cues and improves the accuracy of extracting key fault information. The generated comprehensive vector serves as a knowledge carrier for the maintenance video, enhancing the multi-source knowledge base's ability to understand failure modes in complex scenarios and providing richer and more reliable data support for intelligent DFMEA analysis.

[0074] S107. For each core field data and its corresponding semantic vector, each paragraph in the unstructured document and its corresponding semantic vector, the design drawings and their visual feature vectors and text descriptions, the complaint recordings and their key complaint content and semantic vectors, and the maintenance videos and their key graphic and textual information and comprehensive vectors, respectively, perform association and storage to obtain a multi-source knowledge base.

[0075] In this approach, the processed data and vector representations from each of the aforementioned heterogeneous data sources are stored together in a multi-source knowledge base. This association mechanism ensures that once a vector is matched during the retrieval phase, its corresponding content can be quickly located and retrieved, ultimately forming a multi-source knowledge base that supports multimodal joint queries. Therefore, this construction method achieves the transformation of heterogeneous data into unified, searchable knowledge, providing a solid data foundation for intelligent DFMEA analysis.

[0076] Therefore, the multi-source knowledge base constructed through steps S101-S107 includes a relational database and a vector database. The relational database includes several analytical knowledge items, their source categories, and inclusion times. The analytical knowledge items are categorized as follows: core field data from the DFMEA database, a paragraph from an unstructured document, a design drawing and its text description, a complaint recording and its key complaint content, and a maintenance video and its key graphic information. The vector database includes vectors corresponding to each type of analytical knowledge. Therefore, the relational database supports structured queries and source verification, while the vector database supports nearest neighbor search based on semantic similarity.

[0077] When searching multi-source knowledge bases, relying solely on a single search mode is insufficient to balance terminological precision and semantic generalization, potentially leading to omissions or false positives of key knowledge. To improve the comprehensiveness and accuracy of search results, this invention simultaneously employs keyword and vector search methods to retrieve information from the multi-source knowledge base. The results from both methods are then comprehensively ranked before the final reference analysis knowledge is selected. This achieves efficient filtering and ranking of massive amounts of analytical knowledge from the multi-source knowledge base, ensuring that the contextual information upon which subsequent tasks are based is both relevant and reliable.

[0078] Therefore, in Figure 1 Based on this, please refer to Figure 3 The implementation of step S203 can include S2031 to S2035.

[0079] S2031. Search for each synonym that matches the subject being analyzed from the preset thesaurus.

[0080] In this embodiment, the preset thesaurus includes multiple sets of synonyms, each set containing several synonyms with similar meanings. For example, assuming the subject of analysis is "thermocouple," the synonyms searched could include "temperature probe," "thermoelectric flameout protection device," and "bimetallic thermistor," etc.

[0081] In technical documentation and engineering practice, the same component or concept often has multiple names due to differences in design specifications, usage habits, or expression context. If the retrieval process relies solely on the original terms entered by the user, relevant knowledge may not be effectively recalled due to name differences, thus affecting the completeness of the analysis results. Therefore, expanding the search with synonyms increases the scope of the query and avoids retrieval failures caused by differences in terminology.

[0082] S2032. For each type of analytical knowledge, a preset keyword matching algorithm is used to calculate the matching degree between the analytical knowledge and each subject. Then, the mean value of each matching degree is calculated to obtain the lexical relevance of the analytical knowledge.

[0083] In this embodiment, the subject in step S2032 is the analysis subject or a synonym subject. That is, each subject among the analysis subject and each synonym subject determines the matching degree between itself and each analysis knowledge through a keyword matching algorithm (such as the BM25 algorithm). Then, the average matching degree between each analysis knowledge and each subject is used as the lexical relevance between the analysis knowledge and the analysis subject. This lexical relevance reflects the strength of the association between the analysis knowledge and the analysis subject in the literal meaning dimension.

[0084] In another implementation, keyword retrieval can also be based on SQL (Structured Query Language) queries to achieve precise searches. That is, for any subject (analysis subject or synonymous subject), an SQL query can be generated based on the analysis intent and the subject, and then each relevant analytical knowledge matching the SQL query can be retrieved from the relational database.

[0085] For example, suppose the failure query is "What are the common failure consequences of thermocouples in stoves?", then the analysis subject is "thermocouples in stoves", and the analysis intent is "query failure consequences". Correspondingly, the two parts, "thermocouples in stoves" and "query failure consequences", can be mapped to specific fields in a relational database. The generated SQL query could be: SELECT a.effect FROM dfmea_table a WHERE a.component = 'thermocouple'. It should be noted that this example is merely illustrative and not intended to be limiting.

[0086] After performing a precise query for each subject, it can be assumed that the lexical relevance of all relevant analytical knowledge retrieved is 1, while the lexical relevance of other analytical knowledge in the relational database besides relevant analytical knowledge is 0. Therefore, the lexical relevance of each type of analytical knowledge can also be obtained using precise SQL queries, which can then be used to participate in the comprehensive relevance calculation in the subsequent step S2034.

[0087] S2033. Use a vectorization model to convert failure questions into question vectors, and calculate the cosine distance between the question vectors and the vectors corresponding to each type of analytical knowledge to obtain the semantic relevance of each type of analytical knowledge.

[0088] In this embodiment, semantic relevance measures the proximity of invalid questions and analytical knowledge in the semantic space, and can identify content that has similar meanings but no identical words.

[0089] S2034. The lexical relevance and semantic relevance corresponding to each type of analytical knowledge are weighted and summed to obtain the comprehensive relevance corresponding to each type of analytical knowledge.

[0090] In this embodiment, the comprehensive relevance can be calculated as follows:

[0091] in, Indicates invalid question With the Analytical knowledge The overall correlation between them; This is a weighting factor used to adjust the contributions of semantic relevance and lexical relevance; Indicates invalid question The main body of the analysis and the first Analytical knowledge Lexical relevance between them Indicates invalid question With the Analytical knowledge The semantic relevance between them.

[0092] S2035. Based on the comprehensive relevance, source category, and inclusion time of each type of analytical knowledge, all analytical knowledge is reordered to filter multiple reference analytical knowledge.

[0093] In multi-source knowledge bases, different analytical knowledge may vary significantly in terms of relevance, reliability, and timeliness. If sorting and filtering are based solely on overall relevance, unreliable or outdated information may be introduced, thus affecting the professionalism and accuracy of the subsequently generated results.

[0094] The core concept of this embodiment is to improve the quality and credibility of the final selected reference analysis knowledge by introducing a multi-dimensional scoring mechanism to refine and re-rank the preliminary search results. Specifically, step S2035 can be implemented in ways that include steps a1 to a4.

[0095] Step a1: For each type of analytical knowledge, determine the source authority score that matches the source category of the analytical knowledge from the mapping relationship between multiple source categories and multiple preset scores.

[0096] For example, if the source category is a DFMEA database, its corresponding preset score can be higher, while if the source category is a user's complaint recording, then the corresponding preset score can be set lower.

[0097] Step a2: Input the difference between the time the analytical knowledge was collected and the current time into the preset decay model to obtain the timeliness score corresponding to the analytical knowledge.

[0098] The validity of technical information often diminishes over time; outdated maintenance records or early design specifications may no longer be applicable to the current product status. Therefore, the preset decay model here can be a function that decreases over time (such as exponential decay or linear decay).

[0099] For example, the preset attenuation model is , It is the decay constant (a positive number that determines the rate of decay). Indicates the difference. Indicates the first Analytical knowledge The timeliness rating. It should be noted that this example is for illustrative purposes only and is not intended to be limiting.

[0100] Step a3: Weight the sum of the comprehensive relevance, source authority score, and failure score corresponding to the analytical knowledge to obtain the comprehensive score of the analytical knowledge.

[0101] The weighting coefficients for the comprehensive relevance, source authority score, and invalidity score can be flexibly set, for example, they can be 0.7, 0.2, and 0.1 respectively.

[0102] For each type of analytical knowledge in the multi-source knowledge base, by performing steps a1 to a3 above, a comprehensive score for each type of analytical knowledge can be obtained. This comprehensive score is a unified numerical indicator that integrates key dimensions such as lexical relevance, semantic relevance, information source reliability, and content freshness, thereby achieving a comprehensive evaluation of the referenceability of analytical knowledge.

[0103] Step a4: Arrange all analytical knowledge in descending order of comprehensive score, and use the top K analytical knowledge as reference analytical knowledge.

[0104] In this embodiment, the higher the overall score ranking, the more reliable the analytical knowledge is. For example, K can be 5 or 7.

[0105] After completing the hybrid retrieval and reordering of multi-source knowledge bases, a set of highly relevant reference analysis knowledge can be obtained. However, these knowledge fragments are still discrete information units and need to be integrated into structured professional output through effective reasoning mechanisms. Traditional automated scripts or rule engines are difficult to handle the complex semantic induction and logical deduction required for DFMEA entry generation.

[0106] This invention utilizes the powerful contextual understanding and text generation capabilities of the Large Language Model (LLM) to generate multiple DFMEA entries based on the reference analysis knowledge obtained from retrieval. That is, the implementation of step S204 can include S2041~S2042.

[0107] S2041. Combine the analysis subject, analysis intent, all reference analysis knowledge, and preset task instructions into a Prompt text.

[0108] In this embodiment, the Prompt text is used to instruct the large language model to summarize and generate multiple DFMEA entries related to the potential failure consequences of the subject of analysis based on the input content, and output each DFMEA entry in a preset format (such as JSON format or standard table fields) after providing a brief description of each DFMEA entry.

[0109] S2042. Input the Prompt text into the large language model to obtain the DFMEA analysis results corresponding to the analysis subject.

[0110] After the Prompt text is input into the large language model, the model, based on its inherent language knowledge and reasoning ability, performs in-depth analysis of the input context. The model identifies key statements in various reference analysis knowledge, such as phenomenon descriptions in user complaints, fault diagnosis in maintenance videos, and principle explanations in design manuals, and cross-compares and logically correlates them. Finally, the large language model generates DFMEA analysis results corresponding to the analysis subject. These results are presented in the form of multiple DFMEA entries, each focusing on potential failure consequences and accompanied by a brief explanation based on the chain of evidence. The overall output is organized according to a preset format (such as JSON or standard table fields).

[0111] By constructing context-rich Prompt texts and invoking a large language model, an intelligent transformation from multi-source evidence to structured professional conclusions is achieved. This not only improves the accuracy and traceability of the generated content but also ensures that the output conforms to the specifications of the DFMEA methodology, providing engineers with high-quality decision support.

[0112] For example, assuming the failure problem is "What are the common failure consequences of thermocouples in stoves?", the reference analysis knowledge retrieved from the multi-source knowledge base includes: (1) Key complaint content in the user complaint recording (file name 'complaint_call_345.wav'): "The stove requires holding the switch for a long time to light it, and the fire goes out as soon as you release it." (2) Key words spoken in a repair video (file name: 'Ignition Failure.mp4'): "Look, this head is all burnt black, it's carbon buildup. The thermocouple is not aligned with the spark plug. Just replace it and it will be fine." (3) The description of “low electromotive force” in the Design Manual: “Insufficient contact between the thermocouple and the flame will result in a low electromotive force”; (4) Contents in the DFMEA database: The consequence of the failure mode [slow signal response] is [excessive ignition time].

[0113] Therefore, the Prompt text and the final generated DFMEA analysis results can be shown in the second row and the fourth row of the table below, respectively:

[0114] The examples above are merely illustrative and are not intended to be limiting.

[0115] Among the optional implementation methods, although large language models possess powerful generation capabilities, their output may still have deviations in terms of professionalism, accuracy, and engineering applicability. This can lead to multiple DFMEA entries in the analysis results containing errors, irrelevant information, or duplicate content. Therefore, to ensure the reliability and authority of the analysis results, it is necessary to introduce the judgment of domain experts to form a human-machine collaborative decision-making mechanism. This invention, after each DFMEA analysis result output, establishes a structured feedback loop, transforming the expert review process into accumulable knowledge assets and usable optimization signals, thereby ensuring the quality of the system output and its long-term evolution.

[0116] That is, please see Figure 4 After performing step S204, the method may further include steps S301 to S306.

[0117] S301. Obtain the review comments from verification experts for each DFMEA item.

[0118] After performing step S204 above, the DFMEA analysis results can be presented to experts for review through a visualization interface. Experts can provide review comments for each DFMEA item on this visualization interface, and then the system triggers differentiated processing procedures based on the different types of review comments.

[0119] S302. If the review opinion for any DFMEA entry is to directly adopt it, then the current verification time and the DFMEA entry are stored in the DFMEA database.

[0120] In this embodiment, if the review opinion for a DFMEA entry is "directly adopted," it indicates that the entry's content is accurate and meets the specification requirements. In this case, the current verification time and the DFMEA entry can be stored in the DFMEA database. This operation achieves formal archiving of high-quality knowledge, making the entry reusable standard content in the DFMEA database and continuously enriching the enterprise's structured knowledge reserves.

[0121] S303. If the review opinion for any DFMEA entry is that the modification is adopted, then the current verification time and the modified DFMEA entry are stored in the DFMEA database, and the Prompt text, the DFMEA entry and the modified DFMEA entry are combined into a sample data and stored in the sample database.

[0122] In this embodiment, if the review opinion for a DFMEA entry is "adopted after modification," it indicates that while the original generated content has basic value, it needs to be corrected. In this case, the current verification time and the DFMEA entry modified by the expert can be stored in the DFMEA database to ensure the accuracy of the stored knowledge.

[0123] At the same time, the Prompt text, the DFMEA entry, and the modified DFMEA entry can be combined into a sample dataset and stored in the sample library. This sample dataset records the complete transformation process from a specific input prompt to an incorrect or incomplete output and then to a correct output. The sample dataset is in the form of a triple of "input -> old output -> new output", so it can serve as an ideal training sample for contrastive learning.

[0124] S304. If the review opinion of any DFMEA entry is "rejection", then obtain the reason for rejection of the DFMEA entry and store the DFMEA entry and the reason for rejection in the optimization library.

[0125] In this embodiment, the content of the optimization library can be used to guide technicians in adjusting the prompt words required by the large language model to optimize the front-end prompt engineering design. For example, if there are many entries rejected due to "duplicate content," technicians can add a restriction to the preset task instructions: if there is duplicate analysis knowledge, please integrate it and do not output duplicate DFMEA entries.

[0126] S305. From the DFMEA analysis results, delete each DFMEA entry with the review comment "rejected", and replace each DFMEA entry with the corresponding modified DFMEA entry, to obtain the target DFMEA analysis results.

[0127] Finally, the original DFMEA analysis results were reconstructed based on all review comments: each DFMEA entry with a review comment of rejection was deleted, and each DFMEA entry with a review comment of adoption after modification was replaced with the corresponding modified DFMEA entry, thereby obtaining the target DFMEA analysis results corresponding to the failure issues.

[0128] S306. If the number of samples in the sample library reaches the preset value, then the large language model is post-trained based on all the sample data in the sample library using a contrastive learning approach.

[0129] In this embodiment, when the number of samples in the sample library reaches a preset value (such as 100), the large language model can be post-trained based on these sample data using a comparative learning approach, making the output of the large language model more rigorous and refined, thus achieving a closed-loop improvement in model generation capabilities.

[0130] After completing a comprehensive DFMEA analysis of the product under test, engineers need to systematically integrate the results scattered across multiple analysis subjects to form a comprehensive analysis report with a complete structure and clear logic. If this process relies on manual compilation, it is not only time-consuming and labor-intensive, but also prone to affecting report quality due to inconsistent formats or unclear hierarchical relationships. To improve delivery efficiency and professionalism, this invention can utilize the natural language organization capabilities of a large language model to automatically perform the summarization and structured arrangement of cross-level content. That is, the method may further include S401~S402: S401. Obtain the target DFMEA analysis results corresponding to multiple analysis subjects of the product under test; It is understandable that DFMEA analysis of the product under test needs to involve different levels and aspects. Therefore, engineers can pre-list multiple failure issues, and then obtain the corresponding target DFMEA analysis results for each failure issue by executing the steps S201~S204 and S301~S305 described above. These target DFMEA analysis results are final entries that have been reviewed and corrected by validation experts, covering analysis topics at different levels. There is a certain hierarchical relationship between these analysis subjects, such as from the system level, component level, to the part level.

[0131] For example, a "gas stove" is a system-level entity, which includes component-level entities such as an "ignition system" and a "gas supply system." The "ignition system" can be further subdivided into component-level entities such as "piezoelectric ceramics," "wires," and "thermocouples." It should be noted that this example is merely illustrative and is not intended to be limiting.

[0132] S402. Based on preset prompts and the target DFMEA analysis results corresponding to multiple analysis subjects, call the large language model to generate a DFMEA analysis report for the product under test.

[0133] In this embodiment, preset prompts can be used to instruct the large language model to integrate and analyze the DFMEA analysis results of each target according to the hierarchical relationship of multiple analysis subjects, so as to generate a DFMEA analysis report. For example, the model can first cluster the results at the component level and classify them into the corresponding component-level modules; then summarize common risks and key issues at the component level; and finally perform an overall evaluation and priority ranking at the system level.

[0134] In this way, large language models can organize the originally discrete analysis results into formal documents with standardized formats according to a preset logical framework.

[0135] This invention also provides a computer program product that, when executed by a processor, implements the DFMEA analysis method disclosed in the above embodiments.

[0136] In summary, this invention provides a DFMEA analysis method and computer program product. After obtaining failure queries related to the product under test, the analysis subject and analysis intent are first determined based on the failure queries. Based on the analysis subject and failure queries, various reference analysis knowledge is retrieved from a pre-built multi-source knowledge base using keyword query and vector query methods, respectively. Finally, based on the analysis subject, analysis intent, and all reference analysis knowledge, a large language model is invoked to generate the DFMEA analysis result corresponding to the analysis subject. The multi-source knowledge base of this invention is created based on multiple heterogeneous data sources and includes a large amount of analysis knowledge. Therefore, the keyword query and vector query methods can accurately retrieve reference analysis knowledge from the multi-source knowledge base, and the final DFMEA analysis result is generated by invoking a large language model. This automates the DFMEA analysis, eliminating the need for manual analysis and improving analysis efficiency.

[0137] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A DFMEA analysis method, characterized in that, The method comprises the following steps: acquiring failure questions related to the product to be tested; determining an analysis subject and an analysis intention based on the failure questions; based on the analysis subject and the failure questions, using keyword query and vector query to retrieve a plurality of reference analysis knowledge from a pre-constructed multi-source knowledge base; the multi-source knowledge base is created based on a plurality of heterogeneous data sources, and includes a knowledge base with a large amount of analysis knowledge; based on the analysis subject, the analysis intention, and all the reference analysis knowledge, calling a large language model to generate a DFMEA analysis result corresponding to the analysis subject.

2. The method of claim 1, wherein, The multi-source knowledge base is constructed by the following methods: extracting core field data of each table item from a DFMEA table in a DFMEA database; using a vectorization model to perform vectorization processing on each piece of core field data to obtain a semantic vector corresponding to each piece of core field data; for any unstructured document obtained, based on the single-input limitation of the vectorization model, the unstructured document is divided into a plurality of semantically complete paragraphs, and each paragraph is converted into a corresponding semantic vector using the vectorization model; for any design drawing obtained, a multi-modal model is called to extract visual feature vectors and text descriptions from the design drawing; for any complaint recording obtained, based on the complaint recording, key complaint content related to product failure and its timestamp are extracted, and the key complaint content is converted into a corresponding semantic vector using the vectorization model; for any repair video obtained, based on the repair video, key graphic information is extracted and converted into a corresponding comprehensive vector; each piece of core field data and its corresponding semantic vector, each paragraph in the unstructured document and its corresponding semantic vector, the design drawing and its visual feature vector and text description, the complaint recording and its key complaint content and semantic vector, and the repair video and its key graphic information and comprehensive vector are respectively associated and stored to obtain the multi-source knowledge base.

3. The method of claim 2, wherein, The step of extracting key graphic information from the repair video and converting the key graphic information into a corresponding comprehensive vector comprises: extracting video speaking content and video content description from the repair video; identifying key speaking content related to failure analysis from the video speaking content; based on the timestamp of the key speaking content in the repair video, at least one key frame image is extracted from the repair video; the key graphic information includes the video content description, key speaking content, and all key frame images; using the vectorization model to convert the video content description and the key speaking content into a text content vector, and using the multi-modal model to extract a visual content vector from all key frame images; weighting and fusing the text content vector and the visual content vector to obtain the comprehensive vector.

4. The method of claim 1, wherein, The multi-source knowledge base includes a relational database and a vector database, the relational database includes a plurality of analysis knowledge and its source category and inclusion time, and the vector database includes a vector corresponding to each kind of analysis knowledge. The step of retrieving multiple reference analysis knowledge from a pre-constructed multi-source knowledge base based on the analysis subject and the failure question respectively adopts a keyword query method and a vector query method, including: Finding each synonymous subject matching the analysis subject from a preset synonym library; For each analysis knowledge, a preset keyword matching algorithm is used to calculate the matching degree between the analysis knowledge and each subject, and then the mean of each matching degree is calculated to obtain the lexical relevance of the analysis knowledge; the subject is the analysis subject or the synonymous subject; The failure question is converted into a problem vector using a vectorization model, and the cosine distance between the problem vector and the vector corresponding to each analysis knowledge is calculated to obtain the semantic relevance of each analysis knowledge; The lexical relevance and semantic relevance corresponding to each analysis knowledge are respectively weighted and summed to obtain the comprehensive relevance corresponding to each analysis knowledge; Based on the comprehensive relevance, source category and collection time corresponding to each analysis knowledge, all the analysis knowledge is reordered to filter multiple reference analysis knowledge.

5. The method of claim 4, wherein, The step of reordering all the analysis knowledge based on the comprehensive relevance, source category and collection time corresponding to each analysis knowledge to filter multiple reference analysis knowledge, including: For each analysis knowledge, determine the source authority score matching the source category of the analysis knowledge from the mapping relationship between multiple source categories and multiple preset scores; Input the difference between the collection time of the analysis knowledge and the current time into a preset decay model to obtain the timeliness score corresponding to the analysis knowledge; The comprehensive relevance, source authority score and failure score corresponding to the analysis knowledge are weighted and summed to obtain the comprehensive score of the analysis knowledge; After arranging all the analysis knowledge in descending order of the comprehensive score, the top K analysis knowledge is selected as the reference analysis knowledge.

6. The method of claim 1, wherein, The step of calling a large language model to generate the DFMEA analysis result corresponding to the analysis subject based on the analysis subject, the analysis intent and all the reference analysis knowledge, including: Combine the analysis subject, the analysis intent, all the reference analysis knowledge and a preset task instruction into a Prompt text; Input the Prompt text into the large language model to obtain the DFMEA analysis result corresponding to the analysis subject, wherein the Prompt text is used to instruct the large language model to summarize and generate multiple DFMEA items related to the potential failure consequences of the analysis subject based on the input content, and then output in a preset format after briefly describing each DFMEA item.

7. The method of claim 1, wherein, The DFMEA analysis result includes multiple DFMEA items, and the method further includes: Obtain the review opinion of a verification expert for each DFMEA item; If the review opinion of any DFMEA item is direct adoption, store the current verification time and the DFMEA item into a DFMEA database; If the review opinion of any of the DFMEA entries is to be adopted after modification, the current verification time and the modified DFMEA entry are stored in the DFMEA database, and the Prompt text, the DFMEA entry and the modified DFMEA entry are combined as a sample data to be stored in a sample library; If the review opinion of any of the DFMEA entries is to be rejected, the rejection reason of the DFMEA entry is obtained, and the DFMEA entry and the rejection reason are stored in an optimization library; the optimization library is used to guide the technical personnel to adjust the prompt words; From the DFMEA analysis result, each DFMEA entry with the review opinion of being rejected is deleted, and each DFMEA entry with the review opinion of being adopted after modification is replaced by the corresponding modified DFMEA entry, to obtain a target DFMEA analysis result.

8. The method of claim 7, wherein, The method further comprises: If the number of samples in the sample library reaches a preset value, the big language model is post-trained based on all sample data in the sample library in a comparative learning manner.

9. The method of claim 7, wherein, The method further comprises: Obtaining target DFMEA analysis results corresponding to a plurality of analysis subjects of the to-be-tested product; Based on a preset prompt word and the target DFMEA analysis results corresponding to the plurality of analysis subjects, a big language model is called to generate a DFMEA analysis report of the to-be-tested product; the preset prompt word is used to instruct the big language model to perform integrated analysis on each target DFMEA analysis result according to the hierarchical relationship of the plurality of analysis subjects, so as to generate the DFMEA analysis report.

10. A computer program product, characterised in that, The computer program product, when executed by a processor, implements the method of any one of claims 1-9.

Images